PMF-LP: the first 10 m plastic-mulched farmland distribution map (2019–2021) in the Loess Plateau of China generated using training sample generation and classifier transfer method

Abstract. Plastic film mulching has been extensively used to increase crop yields in arid and semi-arid regions, but it also altered agricultural landscapes and caused severe environmental pollution. Therefore, accurate and timely mapping of plastic-mulched farmland (PMF) distributions is crucial for planning agricultural production and preventing micro-plastic pollution. However, the scarcity of sufficient and representative training samples hinders large-scale supervised classification and extraction of PMF. Additionally, it remains unclear whether a pre-trained classifier can be directly applied to different regions and years for rapid PMF mapping. To address these challenges, we proposed a new framework that simultaneously takes advantages of the automation of index-based method and the generalization ability of supervised classifier-based approach for PMF mapping. Based on the distinctive spectral responses induced by plastic film deployment events, two novel and robust PMF indices—the Max Blue Band-based Plastic-mulched Farmland Index (MBPMFI) and the Blue Band-based Plastic-mulched Farmland Index (BPMFI)—were initially designed to automatically and rapidly extract PMF pixels in cloud-free areas as candidate training samples. Additionally, the transferability of classifiers pre-trained with these automatically generated samples and optimal features was further evaluated in spatial and spatial–temporal transferability scenarios using F1 values. Finally, by coupling the index-based training sample generation method with the temporal classifier transfer approach, PMF distributions were rapidly produced for the Loess Plateau of China (PMF-LP) for 2019–2021. The results showed that the two newly established indices, MBPMFI and BPMFI, were more robust than the existing PMF indices in enhancing PMF information and suppressing complicated backgrounds. The temporal classifier transfer proved suitable for directly and rapidly mapping PMF across multiple years without additional training samples. Using the locally adaptive classifiers as a reference, the average accuracy decrease of the transferred classifiers was less than 7.0 % under the temporal transferability scenario. Our mapping framework achieved F1 values of 0.80–0.86 in recognizing PMF distributions for the Loess Plateau, highlighting its ability to delineate large-scale spatial patterns of PMF. Additionally, the estimated PMF areas based on the PMF-LP aligned well with the agricultural census data at municipal level (R² > 0.87). The framework developed in this study lays a foundation for future monitoring of PMF distributions and agricultural micro-plastic pollution on a large scale. The full archive of PMF-LP is freely available at https://doi.org/10.5281/zenodo.13369426 (Zhao et al., 2024).